Benchmarks
Measured on realsecurity work.
Same 16 tasks, run against a frontier general model and our tuned endpoint. Refusals are counted, not vibes.
| Task family | General model | ZeroDayAPI |
|---|---|---|
| Shellcode offset mapping | refuses ~40% of samples | 4/4 answered |
| PowerShell deobfuscation | refuses on malware heuristic | 4/4 reconstructed |
| OpenAPI payload generation | aborts mid-pipeline | 4/4 executed |
| Ghidra dump triage | partial, ethics preamble | 4/4 classified |
Sample size is small and honest: 16 hand-built tasks, temperature 0.1, October 2026.
Registry
Specialised tiers.
zeroday-offensive-70bLlama-3 fine-tuneexploit analysis, pen testing128K
zeroday-reverse-32bDeepSeek-Coderassembly reversing, patch diffing64K
zeroday-triage-8bMistral fine-tuneSIEM correlation, IOC extraction32K